1 /** vim: set sw=4 ts=4:
4 * @author Adam M. Szalkowski & Sebastian Hack
6 * ILP based spilling & rematerialization
8 * Copyright (C) 2006 Universitaet Karlsruhe
9 * Released under the GPL
32 #include "phiclass_t.h"
37 #include <lpp/lpp_net.h>
38 #include <lpp/lpp_cplex.h>
39 //#include <lc_pset.h>
40 #include <libcore/lc_bitset.h>
44 #include "besched_t.h"
49 #include "bespillremat.h"
51 #include "bepressurestat.h"
53 #include "bechordal_t.h"
59 //#define KEEPALIVE /* keep alive all inserted remats and dump graph with remats */
60 #define COLLECT_REMATS /* enable rematerialization */
61 #define COLLECT_INVERSE_REMATS /* enable placement of inverse remats */
62 #define REMAT_WHILE_LIVE /* only remat values that are live */
63 //#define NO_ENLARGE_L1V3N355 /* do not remat after the death of some operand */
64 //#define EXECFREQ_LOOPDEPH /* compute execution frequency from loop depth only */
65 #define MAY_DIE_AT_REMAT /* allow values to die after a pre remat */
66 #define NO_SINGLE_USE_REMATS /* do not repair schedule */
67 //#define KEEPALIVE_SPILLS
68 //#define KEEPALIVE_RELOADS
69 #define GOODWIN_REDUCTION
70 //#define NO_MEMCOPIES
74 #define LPP_SERVER "i44pc52"
75 #define LPP_SOLVER "cplex"
81 #define LOOP_WEIGHT 12
83 #define ILP_TIMEOUT 120
87 typedef struct _spill_ilp_t {
88 const arch_register_class_t *cls;
90 const be_chordal_env_t *chordal_env;
94 pset *all_possible_remats;
99 set *values; /**< for collecting all definitions of values before running ssa-construction */
104 DEBUG_ONLY(firm_dbg_module_t * dbg);
107 typedef int ilp_var_t;
108 typedef int ilp_cst_t;
110 typedef struct _spill_bb_t {
115 typedef struct _remat_t {
116 const ir_node *op; /**< for copy_irn */
117 const ir_node *value; /**< the value which is being recomputed by this remat */
118 ir_node *proj; /**< not NULL if the above op produces a tuple */
119 int cost; /**< cost of this remat */
120 int inverse; /**< nonzero if this is an inverse remat */
124 * Data to be attached to each IR node. For remats this contains the ilp_var
125 * for this remat and for normal ops this contains the ilp_vars for
126 * reloading each operand
128 typedef struct _op_t {
133 remat_t *remat; /** the remat this op belongs to */
134 int pre; /** 1, if this is a pressure-increasing remat */
138 ir_node *op; /** the operation this live range belongs to */
147 typedef struct _defs_t {
149 ir_node *spills; /**< points to the first spill for this value (linked by link field) */
150 ir_node *remats; /**< points to the first definition for this value (linked by link field) */
153 typedef struct _remat_info_t {
154 const ir_node *irn; /**< the irn to which these remats belong */
155 pset *remats; /**< possible remats for this value */
156 pset *remats_by_operand; /**< remats with this value as operand */
159 typedef struct _keyval_t {
164 typedef struct _spill_t {
174 has_reg_class(const spill_ilp_t * si, const ir_node * irn)
176 return chordal_has_class(si->chordal_env, irn);
181 cmp_remat(const void *a, const void *b)
183 const keyval_t *p = a;
184 const keyval_t *q = b;
185 const remat_t *r = p->val;
186 const remat_t *s = q->val;
190 return !(r == s || r->op == s->op);
194 cmp_remat(const void *a, const void *b)
196 const remat_t *r = a;
197 const remat_t *s = a;
199 return !(r == s || r->op == s->op);
203 cmp_spill(const void *a, const void *b, size_t size)
205 const spill_t *p = a;
206 const spill_t *q = b;
208 // return !(p->irn == q->irn && p->bb == q->bb);
209 return !(p->irn == q->irn);
213 set_find_keyval(set * set, void * key)
218 return set_find(set, &query, sizeof(query), HASH_PTR(key));
222 set_insert_keyval(set * set, void * key, void * val)
228 return set_insert(set, &query, sizeof(query), HASH_PTR(key));
232 set_find_def(set * set, ir_node * value)
237 return set_find(set, &query, sizeof(query), HASH_PTR(value));
241 set_insert_def(set * set, ir_node * value)
248 return set_insert(set, &query, sizeof(query), HASH_PTR(value));
252 set_find_spill(set * set, ir_node * value)
257 return set_find(set, &query, sizeof(query), HASH_PTR(value));
260 #define pset_foreach(s,i) for((i)=pset_first((s)); (i); (i)=pset_next((s)))
261 #define set_foreach(s,i) for((i)=set_first((s)); (i); (i)=set_next((s)))
262 #define foreach_post_remat(s,i) for((i)=next_post_remat((s)); (i); (i)=next_post_remat((i)))
263 #define foreach_pre_remat(si,s,i) for((i)=next_pre_remat((si),(s)); (i); (i)=next_pre_remat((si),(i)))
264 #define sched_foreach_op(s,i) for((i)=sched_next_op((s));!sched_is_end((i));(i)=sched_next_op((i)))
267 cmp_remat_info(const void *a, const void *b, size_t size)
269 const remat_info_t *p = a;
270 const remat_info_t *q = b;
272 return !(p->irn == q->irn);
276 cmp_defs(const void *a, const void *b, size_t size)
281 return !(p->value == q->value);
285 cmp_keyval(const void *a, const void *b, size_t size)
287 const keyval_t *p = a;
288 const keyval_t *q = b;
290 return !(p->key == q->key);
294 execution_frequency(const ir_node * irn)
297 #ifndef EXECFREQ_LOOPDEPH
299 return get_block_execfreq(irn) + FUDGE;
301 return get_block_execfreq(get_nodes_block(irn)) + FUDGE;
305 return exp(get_loop_depth(get_irn_loop(irn)) * log(10)) + FUDGE;
307 return exp(get_loop_depth(get_irn_loop(get_nodes_block(irn))) * log(10)) + FUDGE;
312 get_cost(const spill_ilp_t * si, const ir_node * irn)
314 if(be_is_Spill(irn)) {
316 } else if(be_is_Reload(irn)){
319 return arch_get_op_estimated_cost(si->chordal_env->birg->main_env->arch_env, irn);
325 * Checks, whether node and its operands have suitable reg classes
328 is_rematerializable(const spill_ilp_t * si, const ir_node * irn)
331 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
332 int remat = (arch_irn_get_flags(arch_env, irn) & arch_irn_flags_rematerializable) != 0;
336 ir_fprintf(stderr, " Node %+F is not rematerializable\n", irn);
339 for (n = get_irn_arity(irn)-1; n>=0 && remat; --n) {
340 ir_node *op = get_irn_n(irn, n);
341 remat &= has_reg_class(si, op) || arch_irn_get_flags(arch_env, op) & arch_irn_flags_ignore || (get_irn_op(op) == op_NoMem);
344 // ir_fprintf(stderr, " Argument %d (%+F) of Node %+F has wrong regclass\n", i, op, irn);
351 * Try to create a remat from @p op with destination value @p dest_value
353 static INLINE remat_t *
354 get_remat_from_op(spill_ilp_t * si, const ir_node * dest_value, const ir_node * op)
356 remat_t *remat = NULL;
358 // if(!mode_is_datab(get_irn_mode(dest_value)))
361 if(dest_value == op) {
362 const ir_node *proj = NULL;
364 if(is_Proj(dest_value)) {
365 op = get_irn_n(op, 0);
369 if(!is_rematerializable(si, op))
372 remat = obstack_alloc(si->obst, sizeof(*remat));
374 remat->cost = get_cost(si, op);
375 remat->value = dest_value;
379 arch_inverse_t inverse;
382 /* get the index of the operand we want to retrieve by the inverse op */
383 for (n = get_irn_arity(op)-1; n>=0; --n) {
384 ir_node *arg = get_irn_n(op, n);
386 if(arg == dest_value) break;
390 DBG((si->dbg, LEVEL_5, "\t requesting inverse op for argument %d of op %+F\n", n, op));
392 /* else ask the backend to give an inverse op */
393 if(arch_get_inverse(si->chordal_env->birg->main_env->arch_env, op, n, &inverse, si->obst)) {
396 DBG((si->dbg, LEVEL_4, "\t backend gave us an inverse op with %d nodes and cost %d\n", inverse.n, inverse.costs));
398 assert(inverse.n > 0 && "inverse op should have at least one node");
400 for(i=inverse.n-1; i>=0; --i) {
401 pset_insert_ptr(si->inverse_ops, inverse.nodes[i]);
405 remat = obstack_alloc(si->obst, sizeof(*remat));
406 remat->op = inverse.nodes[0];
407 remat->cost = inverse.costs;
408 remat->value = dest_value;
409 remat->proj = (inverse.n==2)?inverse.nodes[1]:NULL;
412 assert(is_Proj(remat->proj));
414 assert(0 && "I can not handle remats with more than 2 nodes");
421 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F with %+F\n", remat->op, dest_value, op, remat->proj));
423 DBG((si->dbg, LEVEL_3, "\t >Found remat %+F for %+F from %+F\n", remat->op, dest_value, op));
431 add_remat(const spill_ilp_t * si, const remat_t * remat)
433 remat_info_t *remat_info,
438 assert(remat->value);
440 query.irn = remat->value;
442 query.remats_by_operand = NULL;
443 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(remat->value));
445 if(remat_info->remats == NULL) {
446 remat_info->remats = new_pset(cmp_remat, 4096);
448 pset_insert(remat_info->remats, remat, HASH_PTR(remat->op));
450 /* insert the remat into the remats_be_operand set of each argument of the remat op */
451 for (n = get_irn_arity(remat->op)-1; n>=0; --n) {
452 ir_node *arg = get_irn_n(remat->op, n);
456 query.remats_by_operand = NULL;
457 remat_info = set_insert(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
459 if(remat_info->remats_by_operand == NULL) {
460 remat_info->remats_by_operand = new_pset(cmp_remat, 4096);
462 pset_insert(remat_info->remats_by_operand, remat, HASH_PTR(remat->op));
467 get_irn_n_nonremat_edges(const spill_ilp_t * si, const ir_node * irn)
469 const ir_edge_t *edge = get_irn_out_edge_first(irn);
473 if(!pset_find_ptr(si->inverse_ops, edge->src)) {
476 edge = get_irn_out_edge_next(irn, edge);
483 get_remats_from_op(spill_ilp_t * si, const ir_node * op)
488 #ifdef NO_SINGLE_USE_REMATS
489 if(has_reg_class(si, op) && (get_irn_n_nonremat_edges(si, op) > 1)) {
491 if(has_reg_class(si, op)) {
493 remat = get_remat_from_op(si, op, op);
495 add_remat(si, remat);
499 #ifdef COLLECT_INVERSE_REMATS
500 /* repeat the whole stuff for each remat retrieved by get_remat_from_op(op, arg)
502 for (n = get_irn_arity(op)-1; n>=0; --n) {
503 ir_node *arg = get_irn_n(op, n);
505 if(has_reg_class(si, arg)) {
506 /* try to get an inverse remat */
507 remat = get_remat_from_op(si, arg, op);
509 add_remat(si, remat);
518 value_is_defined_before(const spill_ilp_t * si, const ir_node * pos, const ir_node * val)
521 ir_node *def_block = get_nodes_block(val);
527 /* if pos is at end of a basic block */
529 ret = (pos == def_block || block_dominates(def_block, pos));
530 // ir_fprintf(stderr, "(def(bb)=%d) ", ret);
534 /* else if this is a normal operation */
535 block = get_nodes_block(pos);
536 if(block == def_block) {
537 if(!sched_is_scheduled(val)) return 1;
539 ret = sched_comes_after(val, pos);
540 // ir_fprintf(stderr, "(def(same block)=%d) ",ret);
544 ret = block_dominates(def_block, block);
545 // ir_fprintf(stderr, "(def(other block)=%d) ", ret);
549 static INLINE ir_node *
550 sched_block_last_noncf(const spill_ilp_t * si, const ir_node * bb)
552 return sched_skip((ir_node*)bb, 0, sched_skip_cf_predicator, (void *) si->chordal_env->birg->main_env->arch_env);
556 * Returns first non-Phi node of block @p bb
558 static INLINE ir_node *
559 sched_block_first_nonphi(const ir_node * bb)
561 return sched_skip((ir_node*)bb, 1, sched_skip_phi_predicator, NULL);
565 sched_skip_proj_predicator(const ir_node * irn, void * data)
567 return (is_Proj(irn));
570 static INLINE ir_node *
571 sched_next_nonproj(const ir_node * irn, int forward)
573 return sched_skip((ir_node*)irn, forward, sched_skip_proj_predicator, NULL);
577 * Returns next operation node (non-Proj) after @p irn
578 * or the basic block of this node
580 static INLINE ir_node *
581 sched_next_op(const ir_node * irn)
583 ir_node *next = sched_next(irn);
588 return sched_next_nonproj(next, 1);
592 * Returns previous operation node (non-Proj) before @p irn
593 * or the basic block of this node
595 static INLINE ir_node *
596 sched_prev_op(const ir_node * irn)
598 ir_node *prev = sched_prev(irn);
603 return sched_next_nonproj(prev, 0);
607 sched_put_after(ir_node * insert, ir_node * irn)
609 if(is_Block(insert)) {
610 insert = sched_block_first_nonphi(insert);
612 insert = sched_next_op(insert);
614 sched_add_before(insert, irn);
618 sched_put_before(const spill_ilp_t * si, ir_node * insert, ir_node * irn)
620 if(is_Block(insert)) {
621 insert = sched_block_last_noncf(si, insert);
623 insert = sched_next_nonproj(insert, 0);
624 insert = sched_prev(insert);
626 sched_add_after(insert, irn);
630 * Tells you whether a @p remat can be placed before the irn @p pos
633 can_remat_before(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
635 const ir_node *op = remat->op;
641 prev = sched_block_last_noncf(si, pos);
642 prev = sched_next_nonproj(prev, 0);
644 prev = sched_prev_op(pos);
646 /* do not remat if the rematted value is defined immediately before this op */
647 if(prev == remat->op) {
652 /* this should be just fine, the following OP will be using this value, right? */
654 /* only remat AFTER the real definition of a value (?) */
655 if(!value_is_defined_before(si, pos, remat->value)) {
656 // ir_fprintf(stderr, "error(not defined)");
661 for(n=get_irn_arity(op)-1; n>=0 && res; --n) {
662 const ir_node *arg = get_irn_n(op, n);
664 #ifdef NO_ENLARGE_L1V3N355
665 if(has_reg_class(si, arg) && live) {
666 res &= pset_find_ptr(live, arg)?1:0;
668 res &= value_is_defined_before(si, pos, arg);
671 res &= value_is_defined_before(si, pos, arg);
679 * Tells you whether a @p remat can be placed after the irn @p pos
682 can_remat_after(const spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
685 pos = sched_block_first_nonphi(pos);
687 pos = sched_next_op(pos);
690 /* only remat AFTER the real definition of a value (?) */
691 if(!value_is_defined_before(si, pos, remat->value)) {
695 return can_remat_before(si, remat, pos, live);
699 * Collect potetially rematerializable OPs
702 walker_remat_collector(ir_node * irn, void * data)
704 spill_ilp_t *si = data;
706 if(!is_Block(irn) && !is_Phi(irn)) {
707 DBG((si->dbg, LEVEL_4, "\t Processing %+F\n", irn));
708 get_remats_from_op(si, irn);
713 * Inserts a copy of @p irn before @p pos
716 insert_copy_before(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
721 bb = is_Block(pos)?pos:get_nodes_block(pos);
722 copy = exact_copy(irn);
724 _set_phi_class(copy, NULL);
725 set_nodes_block(copy, bb);
726 sched_put_before(si, pos, copy);
732 * Inserts a copy of @p irn after @p pos
735 insert_copy_after(const spill_ilp_t * si, const ir_node * irn, ir_node * pos)
740 bb = is_Block(pos)?pos:get_nodes_block(pos);
741 copy = exact_copy(irn);
743 _set_phi_class(copy, NULL);
744 set_nodes_block(copy, bb);
745 sched_put_after(pos, copy);
751 insert_remat_after(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
755 if(can_remat_after(si, remat, pos, live)) {
760 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
762 copy = insert_copy_after(si, remat->op, pos);
764 ir_snprintf(buf, sizeof(buf), "remat2_%N_%N", copy, pos);
765 op = obstack_alloc(si->obst, sizeof(*op));
767 op->attr.remat.remat = remat;
768 op->attr.remat.pre = 0;
769 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(pos));
771 set_irn_link(copy, op);
772 pset_insert_ptr(si->all_possible_remats, copy);
774 proj_copy = insert_copy_after(si, remat->proj, copy);
775 set_irn_n(proj_copy, 0, copy);
776 set_irn_link(proj_copy, op);
777 pset_insert_ptr(si->all_possible_remats, proj_copy);
789 insert_remat_before(spill_ilp_t * si, const remat_t * remat, const ir_node * pos, const pset * live)
793 if(can_remat_before(si, remat, pos, live)) {
798 DBG((si->dbg, LEVEL_3, "\t >inserting remat %+F\n", remat->op));
800 copy = insert_copy_before(si, remat->op, pos);
802 ir_snprintf(buf, sizeof(buf), "remat_%N_%N", copy, pos);
803 op = obstack_alloc(si->obst, sizeof(*op));
805 op->attr.remat.remat = remat;
806 op->attr.remat.pre = 1;
807 op->attr.remat.ilp = lpp_add_var(si->lpp, buf, lpp_binary, remat->cost*execution_frequency(pos));
809 set_irn_link(copy, op);
810 pset_insert_ptr(si->all_possible_remats, copy);
812 proj_copy = insert_copy_after(si, remat->proj, copy);
813 set_irn_n(proj_copy, 0, copy);
814 set_irn_link(proj_copy, op);
815 pset_insert_ptr(si->all_possible_remats, proj_copy);
827 get_block_n_succs(const ir_node *block) {
828 const ir_edge_t *edge;
830 assert(edges_activated(current_ir_graph));
832 edge = get_block_succ_first(block);
836 edge = get_block_succ_next(block, edge);
841 is_merge_edge(const ir_node * bb)
843 #ifdef GOODWIN_REDUCTION
844 return get_block_n_succs(bb) == 1;
851 is_diverge_edge(const ir_node * bb)
853 #ifdef GOODWIN_REDUCTION
854 return get_Block_n_cfgpreds(bb) == 1;
861 walker_regclass_copy_insertor(ir_node * irn, void * data)
863 spill_ilp_t *si = data;
865 if(is_Phi(irn) && has_reg_class(si, irn)) {
868 for(n=get_irn_arity(irn)-1; n>=0; --n) {
869 ir_node *phi_arg = get_irn_n(irn, n);
870 ir_node *bb = get_Block_cfgpred_block(get_nodes_block(irn), n);
872 if(!has_reg_class(si, phi_arg)) {
873 ir_node *copy = be_new_Copy(si->cls, si->chordal_env->irg, bb, phi_arg);
874 ir_node *pos = sched_block_last_noncf(si, bb);
875 op_t *op = obstack_alloc(si->obst, sizeof(*op));
877 DBG((si->dbg, LEVEL_2, "\t copy to my regclass for arg %+F of %+F\n", phi_arg, irn));
878 sched_add_after(pos, copy);
879 set_irn_n(irn, n, copy);
882 op->attr.live_range.args.reloads = NULL;
883 op->attr.live_range.ilp = ILP_UNDEF;
884 set_irn_link(copy, op);
892 * Insert (so far unused) remats into the irg to
893 * recompute the potential liveness of all values
896 walker_remat_insertor(ir_node * bb, void * data)
898 spill_ilp_t *si = data;
899 spill_bb_t *spill_bb;
903 pset *live = pset_new_ptr_default();
905 DBG((si->dbg, LEVEL_3, "\t Entering %+F\n\n", bb));
907 live_foreach(bb, li) {
908 ir_node *value = (ir_node *) li->irn;
910 /* add remats at end of block */
911 if (live_is_end(li) && has_reg_class(si, value)) {
912 pset_insert_ptr(live, value);
916 spill_bb = obstack_alloc(si->obst, sizeof(*spill_bb));
917 set_irn_link(bb, spill_bb);
919 irn = sched_last(bb);
920 while(!sched_is_end(irn)) {
927 next = sched_prev(irn);
929 DBG((si->dbg, LEVEL_5, "\t at %+F (next: %+F)\n", irn, next));
931 if(is_Phi(irn) || is_Proj(irn)) {
934 if(has_reg_class(si, irn)) {
935 pset_remove_ptr(live, irn);
938 op = obstack_alloc(si->obst, sizeof(*op));
940 op->attr.live_range.args.reloads = NULL;
941 op->attr.live_range.ilp = ILP_UNDEF;
942 set_irn_link(irn, op);
948 op = obstack_alloc(si->obst, sizeof(*op));
950 op->attr.live_range.ilp = ILP_UNDEF;
951 op->attr.live_range.args.reloads = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
952 memset(op->attr.live_range.args.reloads, 0xFF, sizeof(*op->attr.live_range.args.reloads) * get_irn_arity(irn));
953 set_irn_link(irn, op);
955 args = pset_new_ptr_default();
957 /* collect arguments of op */
958 for (n = get_irn_arity(irn)-1; n>=0; --n) {
959 ir_node *arg = get_irn_n(irn, n);
961 pset_insert_ptr(args, arg);
964 /* set args of op already live in epilog */
965 pset_foreach(args, arg) {
966 if(has_reg_class(si, arg)) {
967 pset_insert_ptr(live, arg);
970 /* delete defined value from live set */
971 if(has_reg_class(si, irn)) {
972 pset_remove_ptr(live, irn);
976 remat_args = pset_new_ptr_default();
978 /* insert all possible remats before irn */
979 pset_foreach(args, arg) {
980 remat_info_t *remat_info,
984 /* continue if the operand has the wrong reg class
986 if(!has_reg_class(si, arg))
991 query.remats_by_operand = NULL;
992 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
998 if(remat_info->remats) {
999 pset_foreach(remat_info->remats, remat) {
1000 ir_node *remat_irn = NULL;
1002 DBG((si->dbg, LEVEL_4, "\t considering remat %+F for arg %+F\n", remat->op, arg));
1003 #ifdef REMAT_WHILE_LIVE
1004 if(pset_find_ptr(live, remat->value)) {
1005 remat_irn = insert_remat_before(si, remat, irn, live);
1008 remat_irn = insert_remat_before(si, remat, irn, live);
1011 for(n=get_irn_arity(remat_irn)-1; n>=0; --n) {
1012 ir_node *remat_arg = get_irn_n(remat_irn, n);
1014 if(!has_reg_class(si, remat_arg)) continue;
1016 pset_insert_ptr(remat_args, remat_arg);
1023 /* now we add remat args to op's args because they could also die at this op */
1024 pset_foreach(args,arg) {
1025 if(pset_find_ptr(remat_args, arg)) {
1026 pset_remove_ptr(remat_args, arg);
1029 pset_foreach(remat_args,arg) {
1030 pset_insert_ptr(args, arg);
1033 /* insert all possible remats after irn */
1034 pset_foreach(args, arg) {
1035 remat_info_t *remat_info,
1039 /* continue if the operand has the wrong reg class */
1040 if(!has_reg_class(si, arg))
1044 query.remats = NULL;
1045 query.remats_by_operand = NULL;
1046 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(arg));
1052 /* do not place post remats after jumps */
1053 if(sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) continue;
1055 if(remat_info->remats_by_operand) {
1056 pset_foreach(remat_info->remats_by_operand, remat) {
1057 /* do not insert remats producing the same value as one of the operands */
1058 if(!pset_find_ptr(args, remat->value)) {
1059 DBG((si->dbg, LEVEL_4, "\t considering remat %+F with arg %+F\n", remat->op, arg));
1060 #ifdef REMAT_WHILE_LIVE
1061 if(pset_find_ptr(live, remat->value)) {
1062 insert_remat_after(si, remat, irn, live);
1065 insert_remat_after(si, remat, irn, live);
1072 del_pset(remat_args);
1077 live_foreach(bb, li) {
1078 ir_node *value = (ir_node *) li->irn;
1080 /* add remats at end if successor has multiple predecessors */
1081 if(is_merge_edge(bb)) {
1082 /* add remats at end of block */
1083 if (live_is_end(li) && has_reg_class(si, value)) {
1084 remat_info_t *remat_info,
1089 query.remats = NULL;
1090 query.remats_by_operand = NULL;
1091 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1093 if(remat_info && remat_info->remats) {
1094 pset_foreach(remat_info->remats, remat) {
1095 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at end of block %+F\n", remat->op, bb));
1097 insert_remat_before(si, remat, bb, NULL);
1102 if(is_diverge_edge(bb)) {
1103 /* add remat2s at beginning of block */
1104 if ((live_is_in(li) || (is_Phi(value) && get_nodes_block(value)==bb)) && has_reg_class(si, value)) {
1105 remat_info_t *remat_info,
1110 query.remats = NULL;
1111 query.remats_by_operand = NULL;
1112 remat_info = set_find(si->remat_info, &query, sizeof(query), HASH_PTR(value));
1114 if(remat_info && remat_info->remats) {
1115 pset_foreach(remat_info->remats, remat) {
1116 DBG((si->dbg, LEVEL_4, "\t considering remat %+F at beginning of block %+F\n", remat->op, bb));
1118 /* put the remat here if all its args are available */
1119 insert_remat_after(si, remat, bb, NULL);
1129 * Preparation of blocks' ends for Luke Blockwalker(tm)(R)
1132 luke_endwalker(ir_node * bb, void * data)
1134 spill_ilp_t *si = (spill_ilp_t*)data;
1141 spill_bb_t *spill_bb = get_irn_link(bb);
1144 live = pset_new_ptr_default();
1145 use_end = pset_new_ptr_default();
1147 live_foreach(bb, li) {
1148 irn = (ir_node *) li->irn;
1149 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1152 pset_insert_ptr(live, irn);
1153 op = get_irn_link(irn);
1154 assert(!op->is_remat);
1158 /* collect values used by cond jumps etc. at bb end (use_end) -> always live */
1159 /* their reg_out must always be set */
1160 sched_foreach_reverse(bb, irn) {
1163 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1165 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1166 ir_node *irn_arg = get_irn_n(irn, n);
1168 if(has_reg_class(si, irn_arg)) {
1169 pset_insert_ptr(use_end, irn_arg);
1174 ir_snprintf(buf, sizeof(buf), "check_end_%N", bb);
1175 //cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1176 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - pset_count(use_end));
1178 spill_bb->ilp = new_set(cmp_spill, pset_count(live)+pset_count(use_end));
1180 /* if this is a merge edge we can reload at the end of this block */
1181 if(is_merge_edge(bb)) {
1182 spill_bb->reloads = new_set(cmp_keyval, pset_count(live)+pset_count(use_end));
1183 } else if(pset_count(use_end)){
1184 spill_bb->reloads = new_set(cmp_keyval, pset_count(use_end));
1186 spill_bb->reloads = NULL;
1189 pset_foreach(live,irn) {
1195 /* handle values used by control flow nodes later separately */
1196 if(pset_find_ptr(use_end, irn)) continue;
1199 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1201 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(bb);
1203 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1204 spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1205 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1207 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1208 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1210 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1211 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1213 if(is_merge_edge(bb)) {
1217 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1218 reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(bb));
1219 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1221 /* reload <= mem_out */
1222 rel_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1223 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1224 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1227 spill->reg_in = ILP_UNDEF;
1228 spill->mem_in = ILP_UNDEF;
1231 pset_foreach(use_end,irn) {
1235 ilp_cst_t end_use_req,
1240 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1242 spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(bb);
1244 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", irn, bb);
1245 spill->reg_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1246 /* if irn is used at the end of the block, then it is live anyway */
1247 //lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1249 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1250 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1252 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1253 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1255 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", bb, irn);
1256 reload = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(bb));
1257 set_insert_keyval(spill_bb->reloads, irn, INT_TO_PTR(reload));
1259 /* reload <= mem_out */
1260 rel_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1261 lpp_set_factor_fast(si->lpp, rel_cst, reload, 1.0);
1262 lpp_set_factor_fast(si->lpp, rel_cst, spill->mem_out, -1.0);
1264 spill->reg_in = ILP_UNDEF;
1265 spill->mem_in = ILP_UNDEF;
1267 ir_snprintf(buf, sizeof(buf), "req_cf_end_%N_%N", irn, bb);
1268 end_use_req = lpp_add_cst(si->lpp, buf, lpp_equal, 1);
1269 lpp_set_factor_fast(si->lpp, end_use_req, spill->reg_out, 1.0);
1277 next_post_remat(const ir_node * irn)
1282 irn = sched_block_first_nonphi(irn);
1284 irn = sched_next_op(irn);
1287 if(sched_is_end(irn))
1290 op = (op_t*)get_irn_link(irn);
1291 if(op->is_remat && !op->attr.remat.pre) {
1300 next_pre_remat(const spill_ilp_t * si, const ir_node * irn)
1306 ret = sched_block_last_noncf(si, irn);
1307 ret = sched_next(ret);
1308 ret = sched_prev_op(ret);
1310 ret = sched_prev_op(irn);
1313 if(sched_is_end(ret) || is_Phi(ret))
1316 op = (op_t*)get_irn_link(ret);
1317 if(op->is_remat && op->attr.remat.pre) {
1325 * Find a remat of value @p value in the epilog of @p pos
1328 find_post_remat(const ir_node * value, const ir_node * pos)
1330 while((pos = next_post_remat(pos)) != NULL) {
1333 op = get_irn_link(pos);
1334 assert(op->is_remat && !op->attr.remat.pre);
1336 if(op->attr.remat.remat->value == value)
1337 return (ir_node*)pos;
1340 const ir_edge_t *edge;
1341 foreach_out_edge(pos, edge) {
1342 ir_node *proj = get_edge_src_irn(edge);
1343 assert(is_Proj(proj));
1353 add_to_spill_bb(spill_ilp_t * si, ir_node * bb, ir_node * irn)
1355 spill_bb_t *spill_bb = get_irn_link(bb);
1361 spill = set_find(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1363 double spill_cost = is_Unknown(irn)?0.0001:COST_STORE*execution_frequency(bb);
1365 spill = set_insert(spill_bb->ilp, &query, sizeof(query), HASH_PTR(irn));
1367 spill->reg_out = ILP_UNDEF;
1368 spill->reg_in = ILP_UNDEF;
1369 spill->mem_in = ILP_UNDEF;
1371 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", irn, bb);
1372 spill->mem_out = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1374 ir_snprintf(buf, sizeof(buf), "spill_%N_%N", irn, bb);
1375 spill->spill = lpp_add_var(si->lpp, buf, lpp_binary, spill_cost);
1382 get_live_end(spill_ilp_t * si, ir_node * bb, pset * live)
1387 live_foreach(bb, li) {
1388 irn = (ir_node *) li->irn;
1390 if (live_is_end(li) && has_reg_class(si, irn) && !pset_find_ptr(si->all_possible_remats, irn)) {
1391 pset_insert_ptr(live, irn);
1395 irn = sched_last(bb);
1397 /* all values eaten by control flow operations are also live until the end of the block */
1398 sched_foreach_reverse(bb, irn) {
1401 if(!sched_skip_cf_predicator(irn, si->chordal_env->birg->main_env->arch_env)) break;
1403 for(i=get_irn_arity(irn)-1; i>=0; --i) {
1404 ir_node *arg = get_irn_n(irn,i);
1406 if(has_reg_class(si, arg)) {
1407 pset_insert_ptr(live, arg);
1414 * Inserts ILP-constraints and variables for memory copying before the given position
1417 insert_mem_copy_position(spill_ilp_t * si, pset * live, const ir_node * block)
1419 const ir_node *succ;
1420 const ir_edge_t *edge;
1421 spill_bb_t *spill_bb = get_irn_link(block);
1430 assert(edges_activated(current_ir_graph));
1432 edge = get_block_succ_first(block);
1438 edge = get_block_succ_next(block, edge);
1439 /* next block can only contain phis, if this is a merge edge */
1442 ir_snprintf(buf, sizeof(buf), "copyreg_%N", block);
1443 copyreg = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1445 ir_snprintf(buf, sizeof(buf), "check_copyreg_%N", block);
1446 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
1448 pset_foreach(live, tmp) {
1451 op_t *op = get_irn_link(irn);
1452 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
1454 spill = set_find_spill(spill_bb->ilp, tmp);
1457 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1459 lpp_set_factor_fast(si->lpp, cst, copyreg, 1.0);
1461 sched_foreach(succ, phi) {
1462 const ir_node *to_copy;
1464 spill_t *to_copy_spill;
1465 op_t *phi_op = get_irn_link(phi);
1466 ilp_var_t reload = ILP_UNDEF;
1469 if(!is_Phi(phi)) break;
1470 if(!has_reg_class(si, phi)) continue;
1472 to_copy = get_irn_n(phi, pos);
1474 to_copy_op = get_irn_link(to_copy);
1476 to_copy_spill = set_find_spill(spill_bb->ilp, to_copy);
1477 assert(to_copy_spill);
1479 if(spill_bb->reloads) {
1480 keyval_t *keyval = set_find_keyval(spill_bb->reloads, to_copy);
1483 reload = PTR_TO_INT(keyval->val);
1487 ir_snprintf(buf, sizeof(buf), "req_copy_%N_%N", block, to_copy);
1488 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1490 /* copy - reg_out - reload - remat - live_range <= 0 */
1491 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1492 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1493 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1494 lpp_set_factor_fast(si->lpp, cst, to_copy_op->attr.live_range.ilp, -1.0);
1495 foreach_pre_remat(si, block, tmp) {
1496 op_t *remat_op = get_irn_link(tmp);
1497 if(remat_op->attr.remat.remat->value == to_copy) {
1498 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1502 ir_snprintf(buf, sizeof(buf), "copyreg_%N_%N", block, to_copy);
1503 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1505 /* copy - reg_out - copyreg <= 0 */
1506 lpp_set_factor_fast(si->lpp, cst, phi_op->attr.live_range.args.copies[pos], 1.0);
1507 lpp_set_factor_fast(si->lpp, cst, to_copy_spill->reg_out, -1.0);
1508 lpp_set_factor_fast(si->lpp, cst, copyreg, -1.0);
1514 * Walk all irg blocks and emit this ILP
1517 luke_blockwalker(ir_node * bb, void * data)
1519 spill_ilp_t *si = (spill_ilp_t*)data;
1524 spill_bb_t *spill_bb = get_irn_link(bb);
1527 pset *defs = pset_new_ptr_default();
1530 live = pset_new_ptr_default();
1532 /****************************************
1533 * B A S I C B L O C K E N D
1534 ***************************************/
1537 /* init live values at end of block */
1538 get_live_end(si, bb, live);
1540 pset_foreach(live, irn) {
1542 ilp_var_t reload = ILP_UNDEF;
1544 spill = set_find_spill(spill_bb->ilp, irn);
1547 if(spill_bb->reloads) {
1548 keyval_t *keyval = set_find_keyval(spill_bb->reloads, irn);
1551 reload = PTR_TO_INT(keyval->val);
1555 op = get_irn_link(irn);
1556 assert(!op->is_remat);
1558 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", irn, bb);
1559 op->attr.live_range.ilp = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1560 op->attr.live_range.op = bb;
1562 ir_snprintf(buf, sizeof(buf), "reg_out_%N_%N", bb, irn);
1563 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1565 /* reg_out - reload - remat - live_range <= 0 */
1566 lpp_set_factor_fast(si->lpp, cst, spill->reg_out, 1.0);
1567 if(reload != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1568 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, -1.0);
1569 foreach_pre_remat(si, bb, tmp) {
1570 op_t *remat_op = get_irn_link(tmp);
1571 if(remat_op->attr.remat.remat->value == irn) {
1572 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1575 /* maybe we should also assure that reg_out >= live_range etc. */
1578 #ifndef NO_MEMCOPIES
1579 insert_mem_copy_position(si, live, bb);
1583 * start new live ranges for values used by remats at end of block
1584 * and assure the remat args are available
1586 foreach_pre_remat(si, bb, tmp) {
1587 op_t *remat_op = get_irn_link(tmp);
1590 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1591 ir_node *remat_arg = get_irn_n(tmp, n);
1592 op_t *arg_op = get_irn_link(remat_arg);
1595 if(!has_reg_class(si, remat_arg)) continue;
1597 /* if value is becoming live through use by remat */
1598 if(!pset_find_ptr(live, remat_arg)) {
1599 ir_snprintf(buf, sizeof(buf), "lr_%N_end%N", remat_arg, bb);
1600 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1602 arg_op->attr.live_range.ilp = prev_lr;
1603 arg_op->attr.live_range.op = bb;
1605 DBG((si->dbg, LEVEL_4, " value %+F becoming live through use by remat at end of block %+F\n", remat_arg, tmp));
1607 pset_insert_ptr(live, remat_arg);
1608 add_to_spill_bb(si, bb, remat_arg);
1611 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
1612 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
1613 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1615 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1616 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1618 /* use reload placed for this argument */
1619 if(spill_bb->reloads) {
1620 keyval_t *keyval = set_find_keyval(spill_bb->reloads, remat_arg);
1623 ilp_var_t reload = PTR_TO_INT(keyval->val);
1625 lpp_set_factor_fast(si->lpp, cst, reload, -1.0);
1630 DBG((si->dbg, LEVEL_4, "\t %d values live at end of block %+F\n", pset_count(live), bb));
1635 /**************************************
1636 * B A S I C B L O C K B O D Y
1637 **************************************/
1639 sched_foreach_reverse_from(sched_block_last_noncf(si, bb), irn) {
1645 ilp_cst_t check_pre,
1652 /* iterate only until first phi */
1656 op = get_irn_link(irn);
1658 if(op->is_remat) continue;
1659 DBG((si->dbg, LEVEL_4, "\t at node %+F\n", irn));
1661 /* collect defined values */
1662 if(has_reg_class(si, irn)) {
1663 pset_insert_ptr(defs, irn);
1667 if(is_Proj(irn)) continue;
1670 * init set of irn's arguments
1671 * and all possibly used values around this op
1672 * and values defined by post remats
1674 args = new_set(cmp_keyval, get_irn_arity(irn));
1675 used = pset_new_ptr(pset_count(live) + get_irn_arity(irn));
1676 remat_defs = pset_new_ptr(pset_count(live));
1678 for (n=get_irn_arity(irn)-1; n>=0; --n) {
1679 ir_node *irn_arg = get_irn_n(irn, n);
1680 if(has_reg_class(si, irn_arg)) {
1681 set_insert_keyval(args, irn_arg, (void*)n);
1682 pset_insert_ptr(used, irn_arg);
1685 foreach_post_remat(irn, tmp) {
1686 op_t *remat_op = get_irn_link(tmp);
1688 pset_insert_ptr(remat_defs, remat_op->attr.remat.remat->value);
1690 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1691 ir_node *remat_arg = get_irn_n(tmp, n);
1692 if(has_reg_class(si, remat_arg)) {
1693 pset_insert_ptr(used, remat_arg);
1697 foreach_pre_remat(si, irn, tmp) {
1698 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1699 ir_node *remat_arg = get_irn_n(tmp, n);
1700 if(has_reg_class(si, remat_arg)) {
1701 pset_insert_ptr(used, remat_arg);
1706 /**********************************
1707 * I N E P I L O G O F irn
1708 **********************************/
1710 /* ensure each dying value is used by only one post remat */
1711 pset_foreach(live, tmp) {
1712 ir_node *value = tmp;
1713 op_t *value_op = get_irn_link(value);
1718 foreach_post_remat(irn, remat) {
1719 op_t *remat_op = get_irn_link(remat);
1721 for(n=get_irn_arity(remat)-1; n>=0; --n) {
1722 ir_node *remat_arg = get_irn_n(remat, n);
1724 /* if value is used by this remat add it to constraint */
1725 if(remat_arg == value) {
1727 /* sum remat2s <= 1 + n_remats*live_range */
1728 ir_snprintf(buf, sizeof(buf), "dying_lr_%N_%N", value, irn);
1729 cst = lpp_add_cst(si->lpp, buf, lpp_less, 1.0);
1733 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1739 if(value_op->attr.live_range.ilp != ILP_UNDEF && cst != ILP_UNDEF) {
1740 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, -n_remats);
1746 /* new live ranges for values from L\U defined by post remats */
1747 pset_foreach(live, tmp) {
1748 ir_node *value = tmp;
1749 op_t *value_op = get_irn_link(value);
1751 if(!set_find_keyval(args, value) && !pset_find_ptr(defs, value)) {
1752 ilp_var_t prev_lr = ILP_UNDEF;
1755 if(pset_find_ptr(remat_defs, value)) {
1757 /* next_live_range <= prev_live_range + sum remat2s */
1758 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", value, irn);
1759 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1761 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", value, irn);
1762 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1764 lpp_set_factor_fast(si->lpp, cst, value_op->attr.live_range.ilp, 1.0);
1765 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1767 foreach_post_remat(irn, remat) {
1768 op_t *remat_op = get_irn_link(remat);
1770 /* if value is being rematerialized by this remat */
1771 if(value == remat_op->attr.remat.remat->value) {
1772 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1776 value_op->attr.live_range.ilp = prev_lr;
1777 value_op->attr.live_range.op = irn;
1782 /* requirements for post remats and start live ranges from L/U' for values dying here */
1783 foreach_post_remat(irn, tmp) {
1784 op_t *remat_op = get_irn_link(tmp);
1787 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
1788 ir_node *remat_arg = get_irn_n(tmp, n);
1789 op_t *arg_op = get_irn_link(remat_arg);
1791 if(!has_reg_class(si, remat_arg)) continue;
1793 /* only for values in L\U (TODO and D?), the others are handled with post_use */
1794 if(!pset_find_ptr(used, remat_arg)) {
1795 /* remat <= live_range(remat_arg) */
1796 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_arg_%N", tmp, remat_arg);
1797 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1799 /* if value is becoming live through use by remat2 */
1800 if(!pset_find_ptr(live, remat_arg)) {
1803 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", remat_arg, irn);
1804 lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1806 arg_op->attr.live_range.ilp = lr;
1807 arg_op->attr.live_range.op = irn;
1809 DBG((si->dbg, LEVEL_3, " value %+F becoming live through use by remat2 %+F\n", remat_arg, tmp));
1811 pset_insert_ptr(live, remat_arg);
1812 add_to_spill_bb(si, bb, remat_arg);
1815 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1816 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
1821 d = pset_count(defs);
1822 DBG((si->dbg, LEVEL_4, "\t %+F produces %d values in my register class\n", irn, d));
1824 /* count how many regs irn needs for arguments */
1825 u = set_count(args);
1828 /* check the register pressure in the epilog */
1829 /* sum_{L\U'} lr + sum_{U'} post_use <= k - |D| */
1830 ir_snprintf(buf, sizeof(buf), "check_post_%N", irn);
1831 check_post = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - d);
1833 /* add L\U' to check_post */
1834 pset_foreach(live, tmp) {
1835 if(!pset_find_ptr(used, tmp) && !pset_find_ptr(defs, tmp)) {
1836 /* if a live value is not used by irn */
1837 tmp_op = get_irn_link(tmp);
1838 lpp_set_factor_fast(si->lpp, check_post, tmp_op->attr.live_range.ilp, 1.0);
1842 /***********************************************************
1843 * I T E R A T I O N O V E R U S E S F O R E P I L O G
1844 **********************************************************/
1847 pset_foreach(used, tmp) {
1853 op_t *arg_op = get_irn_link(arg);
1856 spill = add_to_spill_bb(si, bb, arg);
1858 /* new live range for each used value */
1859 ir_snprintf(buf, sizeof(buf), "lr_%N_%N", arg, irn);
1860 prev_lr = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1862 /* the epilog stuff - including post_use, check_post, check_post_remat */
1863 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N", arg, irn);
1864 post_use = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
1866 lpp_set_factor_fast(si->lpp, check_post, post_use, 1.0);
1868 /* arg is live throughout epilog if the next live_range is in a register */
1869 if(pset_find_ptr(live, arg)) {
1870 DBG((si->dbg, LEVEL_3, "\t arg %+F is possibly live in epilog of %+F\n", arg, irn));
1872 /* post_use >= next_lr + remat */
1873 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1874 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1875 lpp_set_factor_fast(si->lpp, cst, post_use, -1.0);
1876 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1880 /* if value is not an arg of op and not possibly defined by post remat
1881 * then it may only die and not become live
1883 if(!set_find_keyval(args, arg)) {
1884 /* post_use <= prev_lr */
1885 ir_snprintf(buf, sizeof(buf), "req_post_use_%N_%N", arg, irn);
1886 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1887 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1888 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1890 if(!pset_find_ptr(remat_defs, arg) && pset_find_ptr(live, arg)) {
1891 /* next_lr <= prev_lr */
1892 ir_snprintf(buf, sizeof(buf), "next_lr_%N_%N", arg, irn);
1893 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1894 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1895 lpp_set_factor_fast(si->lpp, cst, prev_lr, -1.0);
1901 /* forall post remat which use arg add a similar cst */
1902 foreach_post_remat(irn, remat) {
1905 for (n=get_irn_arity(remat)-1; n>=0; --n) {
1906 ir_node *remat_arg = get_irn_n(remat, n);
1907 op_t *remat_op = get_irn_link(remat);
1909 if(remat_arg == arg) {
1910 DBG((si->dbg, LEVEL_3, "\t found remat with arg %+F in epilog of %+F\n", arg, irn));
1912 ir_snprintf(buf, sizeof(buf), "post_use_%N_%N-%d", arg, irn, p++);
1913 cst = lpp_add_cst(si->lpp, buf, lpp_greater, 0.0);
1914 lpp_set_factor_fast(si->lpp, cst, post_use, 1.0);
1915 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
1920 /* new live range begins for each used value */
1921 arg_op->attr.live_range.ilp = prev_lr;
1922 arg_op->attr.live_range.op = irn;
1924 /*if(!pset_find_ptr(live, arg)) {
1925 pset_insert_ptr(live, arg);
1926 add_to_spill_bb(si, bb, arg);
1928 pset_insert_ptr(live, arg);
1932 /* just to be sure */
1933 check_post = ILP_UNDEF;
1942 /* check the register pressure in the prolog */
1943 /* sum_{L\U} lr <= k - |U| */
1944 ir_snprintf(buf, sizeof(buf), "check_pre_%N", irn);
1945 check_pre = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs - u);
1947 /* for the prolog remove defined values from the live set */
1948 pset_foreach(defs, tmp) {
1949 pset_remove_ptr(live, tmp);
1952 /***********************************************************
1953 * I T E R A T I O N O V E R A R G S F O R P R O L O G
1954 **********************************************************/
1957 set_foreach(args, keyval) {
1959 ir_node *arg = keyval->key;
1960 int i = PTR_TO_INT(keyval->val);
1961 op_t *arg_op = get_irn_link(arg);
1963 spill = set_find_spill(spill_bb->ilp, arg);
1966 ir_snprintf(buf, sizeof(buf), "reload_%N_%N", arg, irn);
1967 op->attr.live_range.args.reloads[i] = lpp_add_var(si->lpp, buf, lpp_binary, COST_LOAD*execution_frequency(bb));
1969 /* reload <= mem_out */
1970 ir_snprintf(buf, sizeof(buf), "req_reload_%N_%N", arg, irn);
1971 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
1972 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
1973 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, -1.0);
1975 /* requirement: arg must be in register for use */
1976 /* reload + remat + live_range == 1 */
1977 ir_snprintf(buf, sizeof(buf), "req_%N_%N", irn, arg);
1978 cst = lpp_add_cst(si->lpp, buf, lpp_equal, 1.0);
1980 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, 1.0);
1981 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[i], 1.0);
1982 foreach_pre_remat(si, irn, tmp) {
1983 op_t *remat_op = get_irn_link(tmp);
1984 if(remat_op->attr.remat.remat->value == arg) {
1985 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
1990 /* iterate over L\U */
1991 pset_foreach(live, tmp) {
1992 if(!set_find_keyval(args, tmp)) {
1993 /* if a live value is not used by irn */
1994 tmp_op = get_irn_link(tmp);
1995 lpp_set_factor_fast(si->lpp, check_pre, tmp_op->attr.live_range.ilp, 1.0);
2000 /* requirements for remats */
2001 /* start new live ranges for values used by remats */
2002 foreach_pre_remat(si, irn, tmp) {
2003 op_t *remat_op = get_irn_link(tmp);
2006 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2007 ir_node *remat_arg = get_irn_n(tmp, n);
2008 op_t *arg_op = get_irn_link(remat_arg);
2011 if(!has_reg_class(si, remat_arg)) continue;
2013 /* remat <= live_rang(remat_arg) [ + reload(remat_arg) ] */
2014 ir_snprintf(buf, sizeof(buf), "req_remat_%N_arg_%N", tmp, remat_arg);
2015 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2017 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2018 lpp_set_factor_fast(si->lpp, cst, arg_op->attr.live_range.ilp, -1.0);
2020 /* if remat arg is also used by current op then we can use reload placed for this argument */
2021 if((keyval = set_find_keyval(args, remat_arg)) != NULL) {
2022 int index = (int)keyval->val;
2024 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.args.reloads[index], -1.0);
2032 /*************************
2033 * D O N E W I T H O P
2034 *************************/
2036 DBG((si->dbg, LEVEL_4, "\t %d values live at %+F\n", pset_count(live), irn));
2038 pset_foreach(live, tmp) {
2039 assert(has_reg_class(si, tmp));
2042 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2043 ir_node *arg = get_irn_n(irn, n);
2045 assert(!find_post_remat(arg, irn) && "there should be no post remat for an argument of an op");
2048 del_pset(remat_defs);
2052 defs = pset_new_ptr_default();
2057 /***************************************
2058 * B E G I N N I N G O F B L O C K
2059 ***************************************/
2062 /* we are now at the beginning of the basic block, there are only \Phis in front of us */
2063 DBG((si->dbg, LEVEL_3, "\t %d values live at beginning of block %+F\n", pset_count(live), bb));
2065 pset_foreach(live, irn) {
2066 assert(is_Phi(irn) || get_nodes_block(irn) != bb);
2069 /* construct mem_outs for all values */
2071 set_foreach(spill_bb->ilp, spill) {
2072 ir_snprintf(buf, sizeof(buf), "mem_out_%N_%N", spill->irn, bb);
2073 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2075 lpp_set_factor_fast(si->lpp, cst, spill->mem_out, 1.0);
2076 lpp_set_factor_fast(si->lpp, cst, spill->spill, -1.0);
2078 if(pset_find_ptr(live, spill->irn)) {
2079 DBG((si->dbg, LEVEL_5, "\t %+F live at beginning of block %+F\n", spill->irn, bb));
2081 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N", spill->irn, bb);
2082 spill->mem_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2083 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2085 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
2087 op_t *op = get_irn_link(spill->irn);
2089 /* do we have to copy a phi argument? */
2090 op->attr.live_range.args.copies = obstack_alloc(si->obst, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2091 memset(op->attr.live_range.args.copies, 0xFF, sizeof(*op->attr.live_range.args.copies) * get_irn_arity(spill->irn));
2093 for(n=get_irn_arity(spill->irn)-1; n>=0; --n) {
2094 const ir_node *arg = get_irn_n(spill->irn, n);
2100 /* argument already done? */
2101 if(op->attr.live_range.args.copies[n] != ILP_UNDEF) continue;
2103 /* get sum of execution frequencies of blocks with the same phi argument */
2104 for(m=n; m>=0; --m) {
2105 const ir_node *arg2 = get_irn_n(spill->irn, m);
2108 freq += execution_frequency(get_Block_cfgpred_block(bb, m));
2112 /* copies are not for free */
2113 ir_snprintf(buf, sizeof(buf), "copy_%N_%N", arg, spill->irn);
2114 var = lpp_add_var(si->lpp, buf, lpp_binary, COST_STORE * freq);
2116 for(m=n; m>=0; --m) {
2117 const ir_node *arg2 = get_irn_n(spill->irn, m);
2120 op->attr.live_range.args.copies[m] = var;
2124 /* copy <= mem_in */
2125 ir_snprintf(buf, sizeof(buf), "nocopy_%N_%N", arg, spill->irn);
2126 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2127 lpp_set_factor_fast(si->lpp, cst, var, 1.0);
2128 lpp_set_factor_fast(si->lpp, cst, spill->mem_in, -1.0);
2135 /* L\U is empty at bb start */
2136 /* arg is live throughout epilog if it is reg_in into this block */
2138 /* check the register pressure at the beginning of the block
2141 ir_snprintf(buf, sizeof(buf), "check_start_%N", bb);
2142 cst = lpp_add_cst(si->lpp, buf, lpp_less, si->n_regs);
2144 pset_foreach(live, irn) {
2147 spill = set_find_spill(spill_bb->ilp, irn);
2150 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N", irn, bb);
2151 spill->reg_in = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2153 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, 1.0);
2155 /* spill + mem_in <= 1 */
2156 ir_snprintf(buf, sizeof(buf), "nospill_%N_%N", irn, bb);
2157 nospill = lpp_add_cst(si->lpp, buf, lpp_less, 1);
2159 lpp_set_factor_fast(si->lpp, nospill, spill->mem_in, 1.0);
2160 lpp_set_factor_fast(si->lpp, nospill, spill->spill, 1.0);
2163 foreach_post_remat(bb, irn) {
2164 op_t *remat_op = get_irn_link(irn);
2166 DBG((si->dbg, LEVEL_4, "\t next post remat: %+F\n", irn));
2167 assert(remat_op->is_remat && !remat_op->attr.remat.pre);
2169 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2172 /* forall post remats add requirements */
2173 foreach_post_remat(bb, tmp) {
2176 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2177 ir_node *remat_arg = get_irn_n(tmp, n);
2178 op_t *remat_op = get_irn_link(tmp);
2180 if(!has_reg_class(si, remat_arg)) continue;
2182 spill = set_find_spill(spill_bb->ilp, remat_arg);
2185 /* remat <= reg_in_argument */
2186 ir_snprintf(buf, sizeof(buf), "req_remat2_%N_%N_arg_%N", tmp, bb, remat_arg);
2187 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2188 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2189 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, 1.0);
2193 /* mem_in/reg_in for live_in values, especially phis and their arguments */
2194 pset_foreach(live, irn) {
2198 spill = set_find_spill(spill_bb->ilp, irn);
2199 assert(spill && spill->irn == irn);
2201 if(is_Phi(irn) && get_nodes_block(irn) == bb) {
2202 for (n=get_Phi_n_preds(irn)-1; n>=0; --n) {
2205 ir_node *phi_arg = get_Phi_pred(irn, n);
2206 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2207 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2210 /* although the phi is in the right regclass one or more of
2211 * its arguments can be in a different one or at least to
2214 if(has_reg_class(si, phi_arg)) {
2215 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2216 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2217 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2218 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2220 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2221 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2223 spill_p = set_find_spill(spill_bb_p->ilp, phi_arg);
2226 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2227 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2231 /* else assure the value arrives on all paths in the same resource */
2233 for (n=get_Block_n_cfgpreds(bb)-1; n>=0; --n) {
2236 ir_node *bb_p = get_Block_cfgpred_block(bb, n);
2237 spill_bb_t *spill_bb_p = get_irn_link(bb_p);
2240 ir_snprintf(buf, sizeof(buf), "mem_in_%N_%N-%d", irn, bb, p);
2241 mem_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2242 ir_snprintf(buf, sizeof(buf), "reg_in_%N_%N-%d", irn, bb, p++);
2243 reg_in = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2245 lpp_set_factor_fast(si->lpp, mem_in, spill->mem_in, 1.0);
2246 lpp_set_factor_fast(si->lpp, reg_in, spill->reg_in, 1.0);
2248 spill_p = set_find_spill(spill_bb_p->ilp, irn);
2251 lpp_set_factor_fast(si->lpp, mem_in, spill_p->mem_out, -1.0);
2252 lpp_set_factor_fast(si->lpp, reg_in, spill_p->reg_out, -1.0);
2257 /* first live ranges from reg_ins */
2258 pset_foreach(live, irn) {
2259 op_t *op = get_irn_link(irn);
2261 spill = set_find_spill(spill_bb->ilp, irn);
2262 assert(spill && spill->irn == irn);
2264 ir_snprintf(buf, sizeof(buf), "first_lr_%N_%N", irn, bb);
2265 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2266 lpp_set_factor_fast(si->lpp, cst, op->attr.live_range.ilp, 1.0);
2267 lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2269 foreach_post_remat(bb, tmp) {
2270 op_t *remat_op = get_irn_link(tmp);
2272 if(remat_op->attr.remat.remat->value == irn) {
2273 lpp_set_factor_fast(si->lpp, cst, remat_op->attr.remat.ilp, -1.0);
2278 /* walk forward now and compute constraints for placing spills */
2279 /* this must only be done for values that are not defined in this block */
2280 /* TODO are these values at start of block? if yes, just check whether this is a diverge edge and skip the loop */
2281 pset_foreach(live, irn) {
2283 * if value is defined in this block we can anways place the spill directly after the def
2284 * -> no constraint necessary
2286 if(!is_Phi(irn) && get_nodes_block(irn) == bb) continue;
2289 spill = set_find_spill(spill_bb->ilp, irn);
2292 ir_snprintf(buf, sizeof(buf), "req_spill_%N_%N", irn, bb);
2293 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0.0);
2295 lpp_set_factor_fast(si->lpp, cst, spill->spill, 1.0);
2296 if(is_diverge_edge(bb)) lpp_set_factor_fast(si->lpp, cst, spill->reg_in, -1.0);
2299 sched_foreach_op(bb, tmp) {
2300 op_t *op = get_irn_link(tmp);
2302 if(is_Phi(tmp)) continue;
2303 assert(!is_Proj(tmp));
2306 ir_node *value = op->attr.remat.remat->value;
2309 /* only collect remats up to the first use of a value */
2310 lpp_set_factor_fast(si->lpp, cst, op->attr.remat.ilp, -1.0);
2315 for (n=get_irn_arity(tmp)-1; n>=0; --n) {
2316 ir_node *arg = get_irn_n(tmp, n);
2319 /* if a value is used stop collecting remats */
2325 if(cst == ILP_UNDEF) break;
2333 typedef struct _irnlist_t {
2334 struct list_head list;
2338 typedef struct _interference_t {
2339 struct list_head blocklist;
2345 cmp_interference(const void *a, const void *b, size_t size)
2347 const interference_t *p = a;
2348 const interference_t *q = b;
2350 return !(p->a == q->a && p->b == q->b);
2353 static interference_t *
2354 set_find_interference(set * set, ir_node * a, ir_node * b)
2356 interference_t query;
2358 query.a = (a>b)?a:b;
2359 query.b = (a>b)?b:a;
2361 return set_find(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2364 static interference_t *
2365 set_insert_interference(spill_ilp_t * si, set * set, ir_node * a, ir_node * b, ir_node * bb)
2367 interference_t query,
2369 irnlist_t *list = obstack_alloc(si->obst, sizeof(*list));
2373 result = set_find_interference(set, a, b);
2376 list_add(&list->list, &result->blocklist);
2380 query.a = (a>b)?a:b;
2381 query.b = (a>b)?b:a;
2383 result = set_insert(set, &query, sizeof(query), HASH_PTR(PTR_TO_INT(a) ^ PTR_TO_INT(b)));
2385 INIT_LIST_HEAD(&result->blocklist);
2386 list_add(&list->list, &result->blocklist);
2392 values_interfere_in_block(ir_node * bb, ir_node * a, ir_node * b)
2394 const ir_edge_t *edge;
2396 if(get_nodes_block(a) != bb && get_nodes_block(b) != bb) {
2397 /* both values are live in, so they interfere */
2401 /* ensure a dominates b */
2402 if(value_dominates(b,a)) {
2408 assert(get_nodes_block(b) == bb && "at least b should be defined here in this block");
2411 /* the following code is stolen from bera.c */
2412 if(is_live_end(bb, a))
2415 foreach_out_edge(a, edge) {
2416 const ir_node *user = edge->src;
2417 if(get_nodes_block(user) == bb
2420 && value_dominates(b, user))
2428 * Walk all irg blocks and collect interfering values inside of phi classes
2431 luke_interferencewalker(ir_node * bb, void * data)
2433 spill_ilp_t *si = (spill_ilp_t*)data;
2437 live_foreach(bb, li1) {
2438 ir_node *a = (ir_node *) li1->irn;
2439 op_t *a_op = get_irn_link(a);
2441 if(a_op->is_remat) continue;
2443 /* a is only interesting if it is in my register class and if it is inside a phi class */
2444 if (has_reg_class(si, a) && get_phi_class(a)) {
2445 for(li2=li1->next; li2; li2 = li2->next) {
2446 ir_node *b = (ir_node *) li2->irn;
2447 op_t *b_op = get_irn_link(b);
2449 if(b_op->is_remat) continue;
2451 /* a and b are only interesting if they are in the same phi class */
2452 if(has_reg_class(si, b) && get_phi_class(a) == get_phi_class(b)) {
2453 if(values_interfere_in_block(bb, a, b)) {
2454 DBG((si->dbg, LEVEL_4, "\tvalues interfere in %+F: %+F, %+F\n", bb, a, b));
2455 set_insert_interference(si, si->interferences, a, b, bb);
2463 static unsigned int copy_path_id = 0;
2466 write_copy_path_cst(spill_ilp_t *si, pset * copies, ilp_var_t any_interfere)
2473 ir_snprintf(buf, sizeof(buf), "copy_path-%d", copy_path_id++);
2474 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2476 lpp_set_factor_fast(si->lpp, cst, any_interfere, 1.0);
2478 pset_foreach(copies, ptr) {
2479 copy = PTR_TO_INT(ptr);
2480 lpp_set_factor_fast(si->lpp, cst, copy, -1.0);
2485 * @parameter copies contains a path of copies which lead us to irn
2486 * @parameter visited contains a set of nodes already visited on this path
2489 find_copy_path(spill_ilp_t * si, ir_node * irn, ir_node * target, ilp_var_t any_interfere, pset * copies, pset * visited)
2492 op_t *op = get_irn_link(irn);
2494 if(op->is_remat) return;
2496 pset_insert_ptr(visited, irn);
2501 /* visit all operands */
2502 for(n=get_irn_arity(irn)-1; n>=0; --n) {
2503 ir_node *arg = get_irn_n(irn, n);
2504 ilp_var_t copy = op->attr.live_range.args.copies[n];
2506 if(!has_reg_class(si, arg)) continue;
2509 pset_insert(copies, INT_TO_PTR(copy), copy);
2510 write_copy_path_cst(si, copies, any_interfere);
2511 pset_remove(copies, INT_TO_PTR(copy), copy);
2513 if(!pset_find_ptr(visited, arg)) {
2514 pset_insert(copies, INT_TO_PTR(copy), copy);
2515 find_copy_path(si, arg, target, any_interfere, copies, visited);
2516 pset_remove(copies, INT_TO_PTR(copy), copy);
2522 /* visit all uses which are phis */
2523 foreach_out_edge(irn, edge) {
2524 ir_node *user = edge->src;
2525 int pos = edge->pos;
2526 op_t *op = get_irn_link(user);
2529 if(!is_Phi(user)) continue;
2530 if(!has_reg_class(si, user)) continue;
2532 copy = op->attr.live_range.args.copies[pos];
2534 if(user == target) {
2535 pset_insert(copies, INT_TO_PTR(copy), copy);
2536 write_copy_path_cst(si, copies, any_interfere);
2537 pset_remove(copies, INT_TO_PTR(copy), copy);
2539 if(!pset_find_ptr(visited, user)) {
2540 pset_insert(copies, INT_TO_PTR(copy), copy);
2541 find_copy_path(si, user, target, any_interfere, copies, visited);
2542 pset_remove(copies, INT_TO_PTR(copy), copy);
2547 pset_remove_ptr(visited, irn);
2551 gen_copy_constraints(spill_ilp_t * si, ir_node * a, ir_node * b, ilp_var_t any_interfere)
2553 pset * copies = pset_new_ptr_default();
2554 pset * visited = pset_new_ptr_default();
2556 find_copy_path(si, a, b, any_interfere, copies, visited);
2564 memcopyhandler(spill_ilp_t * si)
2566 interference_t *interference;
2568 /* teste Speicherwerte auf Interferenz */
2570 /* analyze phi classes */
2571 phi_class_compute(si->chordal_env->irg);
2573 DBG((si->dbg, LEVEL_2, "\t calling interferencewalker\n"));
2574 irg_block_walk_graph(si->chordal_env->irg, luke_interferencewalker, NULL, si);
2576 // phi_class_free(si->chordal_env->irg);
2578 /* now lets emit the ILP unequations for the crap */
2579 set_foreach(si->interferences, interference) {
2581 ilp_var_t interfere,
2583 ilp_cst_t any_interfere_cst,
2585 const ir_node *a = interference->a;
2586 const ir_node *b = interference->b;
2588 /* any_interf <= \sum interf */
2589 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N", a, b);
2590 any_interfere_cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2591 any_interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2593 lpp_set_factor_fast(si->lpp, any_interfere_cst, any_interfere, 1.0);
2595 list_for_each_entry(irnlist_t, irnlist, &interference->blocklist, list) {
2596 const ir_node *bb = irnlist->irn;
2597 spill_bb_t *spill_bb = get_irn_link(bb);
2604 spilla = set_find_spill(spill_bb->ilp, a);
2608 spillb = set_find_spill(spill_bb->ilp, b);
2611 /* interfere <-> (mem_in_a or spill_a) and (mem_in_b or spill_b): */
2612 /* 1: mem_in_a + mem_in_b + spill_a + spill_b - interfere <= 1 */
2613 /* 2: - mem_in_a - spill_a + interfere <= 0 */
2614 /* 3: - mem_in_b - spill_b + interfere <= 0 */
2615 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N", bb, a, b);
2616 interfere = lpp_add_var(si->lpp, buf, lpp_binary, 0.0);
2618 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-1", bb, a, b);
2619 cst = lpp_add_cst(si->lpp, buf, lpp_less, 1);
2621 lpp_set_factor_fast(si->lpp, cst, interfere, -1.0);
2622 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, 1.0);
2623 lpp_set_factor_fast(si->lpp, cst, spilla->spill, 1.0);
2624 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, 1.0);
2625 lpp_set_factor_fast(si->lpp, cst, spillb->spill, 1.0);
2627 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-2", bb, a, b);
2628 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2630 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2631 if(spilla->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spilla->mem_in, -1.0);
2632 lpp_set_factor_fast(si->lpp, cst, spilla->spill, -1.0);
2634 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N_%N-3", bb, a, b);
2635 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2637 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2638 if(spillb->mem_in != ILP_UNDEF) lpp_set_factor_fast(si->lpp, cst, spillb->mem_in, -1.0);
2639 lpp_set_factor_fast(si->lpp, cst, spillb->spill, -1.0);
2642 lpp_set_factor_fast(si->lpp, any_interfere_cst, interfere, -1.0);
2644 /* any_interfere >= interf */
2645 ir_snprintf(buf, sizeof(buf), "interfere_%N_%N-%N", a, b, bb);
2646 cst = lpp_add_cst(si->lpp, buf, lpp_less, 0);
2648 lpp_set_factor_fast(si->lpp, cst, interfere, 1.0);
2649 lpp_set_factor_fast(si->lpp, cst, any_interfere, -1.0);
2652 /* now that we know whether the two values interfere in memory we can drop constraints to enforce copies */
2653 gen_copy_constraints(si,a,b,any_interfere);
2661 return fabs(x) < 0.00001;
2665 static int mark_remat_nodes_hook(FILE *F, ir_node *n, ir_node *l)
2667 spill_ilp_t *si = get_irg_link(current_ir_graph);
2669 if(pset_find_ptr(si->all_possible_remats, n)) {
2670 op_t *op = (op_t*)get_irn_link(n);
2671 assert(op && op->is_remat);
2673 if(!op->attr.remat.remat->inverse) {
2674 if(op->attr.remat.pre) {
2675 ir_fprintf(F, "color:red info3:\"remat value: %+F\"", op->attr.remat.remat->value);
2677 ir_fprintf(F, "color:orange info3:\"remat2 value: %+F\"", op->attr.remat.remat->value);
2682 op_t *op = (op_t*)get_irn_link(n);
2683 assert(op && op->is_remat);
2685 if(op->attr.remat.pre) {
2686 ir_fprintf(F, "color:cyan info3:\"remat inverse value: %+F\"", op->attr.remat.remat->value);
2688 ir_fprintf(F, "color:lightcyan info3:\"remat2 inverse value: %+F\"", op->attr.remat.remat->value);
2699 dump_graph_with_remats(ir_graph * irg, const char * suffix)
2701 set_dump_node_vcgattr_hook(mark_remat_nodes_hook);
2702 be_dump(irg, suffix, dump_ir_block_graph_sched);
2703 set_dump_node_vcgattr_hook(NULL);
2708 * Edge hook to dump the schedule edges with annotated register pressure.
2711 sched_pressure_edge_hook(FILE *F, ir_node *irn)
2713 if(sched_is_scheduled(irn) && sched_has_prev(irn)) {
2714 ir_node *prev = sched_prev(irn);
2715 fprintf(F, "edge:{sourcename:\"");
2717 fprintf(F, "\" targetname:\"");
2719 fprintf(F, "\" label:\"%d", (int)get_irn_link(irn));
2720 fprintf(F, "\" color:magenta}\n");
2726 sched_block_attr_hook(FILE *F, ir_node *node, ir_node *local)
2728 if(is_Block(node)) {
2729 fprintf(F, " info3:\"execfreq %g\"", execution_frequency(node));
2736 dump_ir_block_graph_sched_pressure(ir_graph *irg, const char *suffix)
2738 DUMP_NODE_EDGE_FUNC old_edge_hook = get_dump_node_edge_hook();
2740 dump_consts_local(0);
2741 set_dump_node_vcgattr_hook(sched_block_attr_hook);
2742 set_dump_node_edge_hook(sched_pressure_edge_hook);
2743 dump_ir_block_graph(irg, suffix);
2744 set_dump_node_vcgattr_hook(NULL);
2745 set_dump_node_edge_hook(old_edge_hook);
2749 walker_pressure_annotator(ir_node * bb, void * data)
2751 spill_ilp_t *si = data;
2755 pset *live = pset_new_ptr_default();
2758 live_foreach(bb, li) {
2759 irn = (ir_node *) li->irn;
2761 if (live_is_end(li) && has_reg_class(si, irn)) {
2762 pset_insert_ptr(live, irn);
2766 set_irn_link(bb, INT_TO_PTR(pset_count(live)));
2768 sched_foreach_reverse(bb, irn) {
2770 set_irn_link(irn, INT_TO_PTR(pset_count(live)));
2774 if(has_reg_class(si, irn)) {
2775 pset_remove_ptr(live, irn);
2776 if(is_Proj(irn)) ++projs;
2779 if(!is_Proj(irn)) projs = 0;
2781 for (n=get_irn_arity(irn)-1; n>=0; --n) {
2782 ir_node *arg = get_irn_n(irn, n);
2784 if(has_reg_class(si, arg)) pset_insert_ptr(live, arg);
2786 set_irn_link(irn, INT_TO_PTR(pset_count(live)+projs));
2793 dump_pressure_graph(spill_ilp_t * si, const char *suffix)
2795 be_dump(si->chordal_env->irg, suffix, dump_ir_block_graph_sched_pressure);
2800 connect_all_remats_with_keep(spill_ilp_t * si)
2808 n_remats = pset_count(si->all_possible_remats);
2810 ins = obstack_alloc(si->obst, n_remats * sizeof(*ins));
2813 pset_foreach(si->all_possible_remats, irn) {
2818 si->keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_remats, ins);
2820 obstack_free(si->obst, ins);
2826 connect_all_spills_with_keep(spill_ilp_t * si)
2835 n_spills = pset_count(si->spills);
2837 ins = obstack_alloc(si->obst, n_spills * sizeof(*ins));
2840 pset_foreach(si->spills, irn) {
2845 keep = be_new_Keep(si->chordal_env->cls, si->chordal_env->irg, get_irg_end_block(si->chordal_env->irg), n_spills, ins);
2847 obstack_free(si->obst, ins);
2851 /** insert a spill at an arbitrary position */
2852 ir_node *be_spill2(const arch_env_t *arch_env, ir_node *irn, ir_node *insert, ir_node *ctx)
2854 ir_node *bl = is_Block(insert)?insert:get_nodes_block(insert);
2855 ir_graph *irg = get_irn_irg(bl);
2856 ir_node *frame = get_irg_frame(irg);
2860 const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, irn, -1);
2861 const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
2863 spill = be_new_Spill(cls, cls_frame, irg, bl, frame, irn, ctx);
2866 * search the right insertion point. a spill of a phi cannot be put
2867 * directly after the phi, if there are some phis behind the one which
2868 * is spilled. Also, a spill of a Proj must be after all Projs of the
2871 * Here's one special case:
2872 * If the spill is in the start block, the spill must be after the frame
2873 * pointer is set up. This is done by setting insert to the end of the block
2874 * which is its default initialization (see above).
2877 if(bl == get_irg_start_block(irg) && sched_get_time_step(frame) >= sched_get_time_step(insert))
2880 for (next = sched_next(insert); is_Phi(next) || is_Proj(next); next = sched_next(insert))
2883 sched_add_after(insert, spill);
2888 delete_remat(spill_ilp_t * si, ir_node * remat) {
2890 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2892 sched_remove(remat);
2894 /* kill links to operands */
2895 for (n=get_irn_arity(remat)-1; n>=-1; --n) {
2896 set_irn_n(remat, n, bad);
2901 clean_remat_info(spill_ilp_t * si)
2905 remat_info_t *remat_info;
2906 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2908 set_foreach(si->remat_info, remat_info) {
2909 if(!remat_info->remats) continue;
2911 pset_foreach(remat_info->remats, remat)
2913 if(remat->proj && get_irn_n_edges(remat->proj) == 0) {
2914 set_irn_n(remat->proj, -1, bad);
2915 set_irn_n(remat->proj, 0, bad);
2918 if(get_irn_n_edges(remat->op) == 0) {
2919 for (n=get_irn_arity(remat->op)-1; n>=-1; --n) {
2920 set_irn_n(remat->op, n, bad);
2925 if(remat_info->remats) del_pset(remat_info->remats);
2926 if(remat_info->remats_by_operand) del_pset(remat_info->remats_by_operand);
2931 delete_unnecessary_remats(spill_ilp_t * si)
2935 ir_node *bad = get_irg_bad(si->chordal_env->irg);
2938 ir_node *end = get_irg_end(si->chordal_env->irg);
2941 for (n=get_irn_arity(si->keep)-1; n>=0; --n) {
2942 ir_node *keep_arg = get_irn_n(si->keep, n);
2943 op_t *arg_op = get_irn_link(keep_arg);
2946 assert(arg_op->is_remat);
2948 name = si->lpp->vars[arg_op->attr.remat.ilp];
2950 if(is_zero(name->value)) {
2951 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", keep_arg));
2952 /* TODO check whether reload is preferred over remat (could be bug) */
2953 delete_remat(si, keep_arg);
2955 if(!arg_op->attr.remat.remat->inverse) {
2956 if(arg_op->attr.remat.pre) {
2957 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", keep_arg));
2959 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", keep_arg));
2962 if(arg_op->attr.remat.pre) {
2963 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", keep_arg));
2965 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", keep_arg));
2970 set_irn_n(si->keep, n, bad);
2973 for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
2974 ir_node *end_arg = get_End_keepalive(end, i);
2976 if(end_arg != si->keep) {
2977 obstack_grow(si->obst, &end_arg, sizeof(end_arg));
2980 keeps = obstack_finish(si->obst);
2981 set_End_keepalives(end, n-1, keeps);
2982 obstack_free(si->obst, keeps);
2985 DBG((si->dbg, LEVEL_2, "\t no remats to delete (none have been inserted)\n"));
2990 pset_foreach(si->all_possible_remats, remat) {
2991 op_t *remat_op = get_irn_link(remat);
2992 lpp_name_t *name = si->lpp->vars[remat_op->attr.remat.ilp];
2994 if(is_zero(name->value)) {
2995 DBG((si->dbg, LEVEL_3, "\t deleting remat %+F\n", remat));
2996 /* TODO check whether reload is preferred over remat (could be bug) */
2997 delete_remat(si, remat);
2999 if(!remat_op->attr.remat.remat->inverse) {
3000 if(remat_op->attr.remat.pre) {
3001 DBG((si->dbg, LEVEL_2, "\t**remat kept: %+F\n", remat));
3003 DBG((si->dbg, LEVEL_2, "\t%%%%remat2 kept: %+F\n", remat));
3006 if(remat_op->attr.remat.pre) {
3007 DBG((si->dbg, LEVEL_2, "\t**INVERSE remat kept: %+F\n", remat));
3009 DBG((si->dbg, LEVEL_2, "\t%%%%INVERSE remat2 kept: %+F\n", remat));
3018 get_spills_for_value(spill_ilp_t * si, ir_node * value)
3020 pset *spills = pset_new_ptr_default();
3025 defs = set_find_def(si->values, value);
3027 if(defs && defs->spills) {
3028 for(next = defs->spills; next; next = get_irn_link(next)) {
3029 pset_insert_ptr(spills, next);
3037 get_remats_for_value(spill_ilp_t * si, ir_node * value)
3039 pset *remats = pset_new_ptr_default();
3044 pset_insert_ptr(remats, value);
3045 defs = set_find_def(si->values, value);
3047 if(defs && defs->remats) {
3048 for(next = defs->remats; next; next = get_irn_link(next)) {
3049 pset_insert_ptr(remats, next);
3058 * @param before The node after which the spill will be placed in the schedule
3060 /* TODO set context properly */
3062 insert_spill(spill_ilp_t * si, ir_node * irn, ir_node * value, ir_node * before)
3066 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3068 DBG((si->dbg, LEVEL_3, "\t inserting spill for value %+F after %+F\n", irn, before));
3070 spill = be_spill2(arch_env, irn, before, irn);
3072 defs = set_insert_def(si->values, value);
3075 /* enter into the linked list */
3076 set_irn_link(spill, defs->spills);
3077 defs->spills = spill;
3079 #ifdef KEEPALIVE_SPILLS
3080 pset_insert_ptr(si->spills, spill);
3087 * @param before The Phi node which has to be spilled
3090 insert_mem_phi(spill_ilp_t * si, const ir_node * phi)
3096 op_t *op = get_irn_link(phi);
3098 NEW_ARR_A(ir_node*, ins, get_irn_arity(phi));
3100 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3101 ins[n] = si->m_unknown;
3104 mem_phi = new_r_Phi(si->chordal_env->irg, get_nodes_block(phi), get_irn_arity(phi), ins, mode_M);
3106 defs = set_insert_def(si->values, phi);
3109 /* enter into the linked list */
3110 set_irn_link(mem_phi, defs->spills);
3111 defs->spills = mem_phi;
3113 sched_add_after(phi, mem_phi);
3115 #ifdef KEEPALIVE_SPILLS
3116 pset_insert_ptr(si->spills, mem_phi);
3124 * Add remat to list of defs, destroys link field!
3127 insert_remat(spill_ilp_t * si, ir_node * remat)
3130 op_t *remat_op = get_irn_link(remat);
3132 assert(remat_op->is_remat);
3134 defs = set_insert_def(si->values, remat_op->attr.remat.remat->value);
3137 /* enter into the linked list */
3138 set_irn_link(remat, defs->remats);
3139 defs->remats = remat;
3144 * Add reload before operation and add to list of defs
3147 insert_reload(spill_ilp_t * si, const ir_node * value, const ir_node * after)
3152 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3154 DBG((si->dbg, LEVEL_3, "\t inserting reload for value %+F before %+F\n", value, after));
3156 defs = set_find_def(si->values, value);
3158 spill = defs->spills;
3159 assert(spill && "no spill placed before reload");
3161 reload = be_reload(arch_env, si->cls, after, get_irn_mode(value), spill);
3163 /* enter into the linked list */
3164 set_irn_link(reload, defs->remats);
3165 defs->remats = reload;
3171 walker_spill_placer(ir_node * bb, void * data) {
3172 spill_ilp_t *si = (spill_ilp_t*)data;
3174 spill_bb_t *spill_bb = get_irn_link(bb);
3175 pset *spills_to_do = pset_new_ptr_default();
3178 set_foreach(spill_bb->ilp, spill) {
3181 if(is_Phi(spill->irn) && get_nodes_block(spill->irn) == bb) {
3182 name = si->lpp->vars[spill->mem_in];
3183 if(!is_zero(name->value)) {
3186 mem_phi = insert_mem_phi(si, spill->irn);
3188 DBG((si->dbg, LEVEL_2, "\t >>spilled Phi %+F -> %+F\n", spill->irn, mem_phi));
3192 name = si->lpp->vars[spill->spill];
3193 if(!is_zero(name->value)) {
3194 /* place spill directly after definition */
3195 if(get_nodes_block(spill->irn) == bb) {
3196 insert_spill(si, spill->irn, spill->irn, spill->irn);
3200 /* place spill at bb start */
3201 if(spill->reg_in > 0) {
3202 name = si->lpp->vars[spill->reg_in];
3203 if(!is_zero(name->value)) {
3204 insert_spill(si, spill->irn, spill->irn, bb);
3208 /* place spill after a remat */
3209 pset_insert_ptr(spills_to_do, spill->irn);
3212 DBG((si->dbg, LEVEL_3, "\t %d spills to do in block %+F\n", pset_count(spills_to_do), bb));
3215 for(irn = sched_block_first_nonphi(bb); !sched_is_end(irn); irn = sched_next(irn)) {
3216 op_t *op = get_irn_link(irn);
3218 if(be_is_Spill(irn)) continue;
3221 /* TODO fix this if we want to support remats with more than two nodes */
3222 if(get_irn_mode(irn) != mode_T && pset_find_ptr(spills_to_do, op->attr.remat.remat->value)) {
3223 pset_remove_ptr(spills_to_do, op->attr.remat.remat->value);
3225 insert_spill(si, irn, op->attr.remat.remat->value, irn);
3228 if(pset_find_ptr(spills_to_do, irn)) {
3229 pset_remove_ptr(spills_to_do, irn);
3231 insert_spill(si, irn, irn, irn);
3237 assert(pset_count(spills_to_do) == 0);
3239 /* afterwards free data in block */
3240 del_pset(spills_to_do);
3244 insert_mem_copy(spill_ilp_t * si, const ir_node * bb, const ir_node * value)
3246 ir_node *insert_pos = bb;
3248 const arch_env_t *arch_env = si->chordal_env->birg->main_env->arch_env;
3250 /* find last definition of arg value in block */
3255 defs = set_find_def(si->values, value);
3257 if(defs && defs->remats) {
3258 for(next = defs->remats; next; next = get_irn_link(next)) {
3259 if(get_nodes_block(next) == bb && sched_get_time_step(next) > last) {
3260 last = sched_get_time_step(next);
3266 if(get_nodes_block(value) == bb && sched_get_time_step(value) > last) {
3267 last = sched_get_time_step(value);
3271 DBG((si->dbg, LEVEL_2, "\t inserting mem copy for value %+F after %+F\n", value, insert_pos));
3273 spill = be_spill2(arch_env, is_Block(insert_pos)?value:insert_pos, insert_pos, value);
3279 phim_fixer(spill_ilp_t *si) {
3282 set_foreach(si->values, defs) {
3283 const ir_node *phi = defs->value;
3284 op_t *op = get_irn_link(phi);
3285 ir_node *phi_m = NULL;
3286 ir_node *next = defs->spills;
3289 if(!is_Phi(phi)) continue;
3292 if(is_Phi(next) && get_irn_mode(next) == mode_M) {
3296 next = get_irn_link(next);
3299 if(!phi_m) continue;
3301 for(n=get_irn_arity(phi)-1; n>=0; --n) {
3302 const ir_node *value = get_irn_n(phi, n);
3303 defs_t *val_defs = set_find_def(si->values, value);
3304 ir_node *arg = get_irn_n(phi_m, n);
3306 /* get a spill of this value */
3307 ir_node *spill = val_defs->spills;
3310 #ifndef NO_MEMCOPIES
3311 ir_node *pred = get_Block_cfgpred_block(get_nodes_block(phi), n);
3312 lpp_name_t *name = si->lpp->vars[op->attr.live_range.args.copies[n]];
3314 if(!is_zero(name->value)) {
3315 spill = insert_mem_copy(si, pred, value);
3317 assert(spill && "no spill placed before PhiM");
3320 assert(spill && "no spill placed before PhiM");
3322 set_irn_n(phi_m, n, spill);
3328 walker_reload_placer(ir_node * bb, void * data) {
3329 spill_ilp_t *si = (spill_ilp_t*)data;
3331 spill_bb_t *spill_bb = get_irn_link(bb);
3335 /* reloads at end of block */
3336 if(spill_bb->reloads) {
3339 set_foreach(spill_bb->reloads, keyval) {
3340 ir_node *irn = (ir_node*)keyval->key;
3341 ilp_var_t reload = PTR_TO_INT(keyval->val);
3344 name = si->lpp->vars[reload];
3345 if(!is_zero(name->value)) {
3347 ir_node *insert_pos = bb;
3348 ir_node *prev = sched_block_last_noncf(si, bb);
3349 op_t *prev_op = get_irn_link(prev);
3351 while(be_is_Spill(prev)) {
3352 prev = sched_prev(prev);
3355 prev_op = get_irn_link(prev);
3357 /* insert reload before pre-remats */
3358 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3359 && prev_op->is_remat && prev_op->attr.remat.pre) {
3363 prev = sched_prev(prev);
3364 } while(be_is_Spill(prev));
3366 prev_op = get_irn_link(prev);
3370 reload = insert_reload(si, irn, insert_pos);
3372 #ifdef KEEPALIVE_RELOADS
3373 pset_insert_ptr(si->spills, reload);
3379 /* walk and insert more reloads and collect remats */
3380 sched_foreach_reverse(bb, irn) {
3381 op_t *op = get_irn_link(irn);
3383 if(be_is_Reload(irn) || be_is_Spill(irn)) continue;
3384 if(is_Phi(irn)) break;
3387 if(get_irn_mode(irn) != mode_T) {
3388 insert_remat(si, irn);
3393 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3394 ir_node *arg = get_irn_n(irn, n);
3396 if(op->attr.live_range.args.reloads && op->attr.live_range.args.reloads[n] != ILP_UNDEF) {
3399 name = si->lpp->vars[op->attr.live_range.args.reloads[n]];
3400 if(!is_zero(name->value)) {
3402 ir_node *insert_pos = irn;
3403 ir_node *prev = sched_prev(insert_pos);
3406 while(be_is_Spill(prev)) {
3407 prev = sched_prev(prev);
3410 prev_op = get_irn_link(prev);
3412 /* insert reload before pre-remats */
3413 while(!sched_is_end(prev) && !be_is_Reload(prev) && !is_Phi(prev)
3414 && prev_op->is_remat && prev_op->attr.remat.pre) {
3418 prev = sched_prev(prev);
3419 } while(be_is_Spill(prev));
3421 prev_op = get_irn_link(prev);
3425 reload = insert_reload(si, arg, insert_pos);
3427 set_irn_n(irn, n, reload);
3429 #ifdef KEEPALIVE_RELOADS
3430 pset_insert_ptr(si->spills, reload);
3438 del_set(spill_bb->ilp);
3439 if(spill_bb->reloads) del_set(spill_bb->reloads);
3443 walker_collect_used(ir_node * irn, void * data)
3445 lc_bitset_t *used = data;
3447 lc_bitset_set(used, get_irn_idx(irn));
3450 struct kill_helper {
3456 walker_kill_unused(ir_node * bb, void * data)
3458 struct kill_helper *kh = data;
3459 const ir_node *bad = get_irg_bad(get_irn_irg(bb));
3463 for(irn=sched_first(bb); !sched_is_end(irn);) {
3464 ir_node *next = sched_next(irn);
3467 if(!lc_bitset_is_set(kh->used, get_irn_idx(irn))) {
3468 if(be_is_Spill(irn) || be_is_Reload(irn)) {
3469 DBG((kh->si->dbg, LEVEL_1, "\t SUBOPTIMAL! %+F IS UNUSED (cost: %g)\n", irn, get_cost(kh->si, irn)*execution_frequency(bb)));
3471 assert(lpp_get_sol_state(kh->si->lpp) != lpp_optimal && "optimal solution is suboptimal?");
3477 set_nodes_block(irn, bad);
3478 for (n=get_irn_arity(irn)-1; n>=0; --n) {
3479 set_irn_n(irn, n, bad);
3487 kill_all_unused_values_in_schedule(spill_ilp_t * si)
3489 struct kill_helper kh;
3491 kh.used = lc_bitset_malloc(get_irg_last_idx(si->chordal_env->irg));
3494 irg_walk_graph(si->chordal_env->irg, walker_collect_used, NULL, kh.used);
3495 irg_block_walk_graph(si->chordal_env->irg, walker_kill_unused, NULL, &kh);
3497 lc_bitset_free(kh.used);
3501 print_irn_pset(pset * p)
3505 pset_foreach(p, irn) {
3506 ir_printf("%+F\n", irn);
3511 rewire_uses(spill_ilp_t * si)
3513 dom_front_info_t *dfi = be_compute_dominance_frontiers(si->chordal_env->irg);
3515 pset *ignore = pset_new_ptr(1);
3517 pset_insert_ptr(ignore, get_irg_end(si->chordal_env->irg));
3519 /* then fix uses of spills */
3520 set_foreach(si->values, defs) {
3523 ir_node *next = defs->remats;
3526 reloads = pset_new_ptr_default();
3529 if(be_is_Reload(next)) {
3530 pset_insert_ptr(reloads, next);
3534 next = get_irn_link(next);
3537 spills = get_spills_for_value(si, defs->value);
3538 DBG((si->dbg, LEVEL_2, "\t %d remats, %d reloads, and %d spills for value %+F\n", remats, pset_count(reloads), pset_count(spills), defs->value));
3539 if(pset_count(spills) > 1) {
3540 //assert(pset_count(reloads) > 0);
3541 // print_irn_pset(spills);
3542 // print_irn_pset(reloads);
3544 be_ssa_constr_set_ignore(dfi, spills, ignore);
3551 /* first fix uses of remats and reloads */
3552 set_foreach(si->values, defs) {
3554 ir_node *next = defs->remats;
3557 nodes = pset_new_ptr_default();
3558 pset_insert_ptr(nodes, defs->value);
3561 pset_insert_ptr(nodes, next);
3562 next = get_irn_link(next);
3565 if(pset_count(nodes) > 1) {
3566 DBG((si->dbg, LEVEL_4, "\t %d new definitions for value %+F\n", pset_count(nodes)-1, defs->value));
3567 be_ssa_constr_set(dfi, nodes);
3574 // remove_unused_defs(si);
3576 be_free_dominance_frontiers(dfi);
3581 writeback_results(spill_ilp_t * si)
3583 /* walk through the graph and collect all spills, reloads and remats for a value */
3585 si->values = new_set(cmp_defs, 4096);
3587 DBG((si->dbg, LEVEL_1, "Applying results\n"));
3588 delete_unnecessary_remats(si);
3589 si->m_unknown = new_r_Unknown(si->chordal_env->irg, mode_M);
3590 irg_block_walk_graph(si->chordal_env->irg, walker_spill_placer, NULL, si);
3591 irg_block_walk_graph(si->chordal_env->irg, walker_reload_placer, NULL, si);
3594 /* clean the remat info! there are still back-edges leading there! */
3595 clean_remat_info(si);
3599 connect_all_spills_with_keep(si);
3601 del_set(si->values);
3605 get_n_regs(spill_ilp_t * si)
3607 int arch_n_regs = arch_register_class_n_regs(si->cls);
3611 for(i=0; i<arch_n_regs; i++) {
3612 if(!arch_register_type_is(&si->cls->regs[i], ignore)) {
3617 DBG((si->dbg, LEVEL_1, "\tArchitecture has %d free registers in class %s\n", free, si->cls->name));
3622 walker_reload_mover(ir_node * bb, void * data)
3624 spill_ilp_t *si = data;
3627 sched_foreach(bb, tmp) {
3628 if(be_is_Reload(tmp) && has_reg_class(si, tmp)) {
3629 ir_node *reload = tmp;
3632 /* move reload upwards */
3634 int pressure = (int)get_irn_link(reload);
3635 if(pressure < si->n_regs) {
3636 irn = sched_prev(reload);
3637 DBG((si->dbg, LEVEL_5, "regpressure before %+F: %d\n", reload, pressure));
3638 sched_remove(reload);
3639 pressure = (int)get_irn_link(irn);
3641 while(pressure < si->n_regs) {
3642 if(sched_is_end(irn) || (be_is_Reload(irn) && has_reg_class(si, irn))) break;
3644 set_irn_link(irn, INT_TO_PTR(pressure+1));
3645 DBG((si->dbg, LEVEL_5, "new regpressure before %+F: %d\n", irn, pressure+1));
3646 irn = sched_prev(irn);
3648 pressure = (int)get_irn_link(irn);
3651 DBG((si->dbg, LEVEL_3, "putting reload %+F after %+F\n", reload, irn));
3652 sched_put_after(irn, reload);
3659 move_reloads_upward(spill_ilp_t * si)
3661 irg_block_walk_graph(si->chordal_env->irg, walker_reload_mover, NULL, si);
3666 * Walk all irg blocks and check for interfering spills inside of phi classes
3669 luke_meminterferencechecker(ir_node * bb, void * data)
3671 spill_ilp_t *si = (spill_ilp_t*)data;
3675 live_foreach(bb, li1) {
3676 ir_node *a = (ir_node *) li1->irn;
3678 if(!be_is_Spill(a) && (!is_Phi(a) || get_irn_mode(a) != mode_T)) continue;
3680 /* a is only interesting if it is inside a phi class */
3681 if (get_phi_class(a)) {
3682 for(li2=li1->next; li2; li2 = li2->next) {
3683 ir_node *b = (ir_node *) li2->irn;
3685 if(!be_is_Spill(b) && (!is_Phi(b) || get_irn_mode(b) != mode_T)) continue;
3687 /* a and b are only interesting if they are in the same phi class */
3688 if(get_phi_class(a) == get_phi_class(b)) {
3689 if(values_interfere_in_block(bb, a, b)) {
3690 ir_fprintf(stderr, "$$ Spills interfere in %+F: %+F, %+F \t$$\n", bb, a, b);
3699 verify_phiclasses(spill_ilp_t * si)
3701 /* analyze phi classes */
3702 phi_class_compute(si->chordal_env->irg);
3704 DBG((si->dbg, LEVEL_2, "\t calling memory interference checker\n"));
3705 irg_block_walk_graph(si->chordal_env->irg, luke_meminterferencechecker, NULL, si);
3709 walker_spillslotassigner(ir_node * irn, void * data)
3711 spill_ilp_t *si = (spill_ilp_t*)data;
3714 if(!be_is_Spill(irn)) return;
3716 /* set spill context to phi class if it has one ;) */
3718 cls = get_phi_class(irn);
3720 be_set_Spill_context(irn, cls);
3722 be_set_Spill_context(irn, irn);
3727 assign_spillslots(spill_ilp_t * si)
3729 DBG((si->dbg, LEVEL_2, "\t calling spill slot assigner\n"));
3730 irg_walk_graph(si->chordal_env->irg, walker_spillslotassigner, NULL, si);
3734 be_spill_remat(const be_chordal_env_t * chordal_env)
3736 char problem_name[256];
3737 char dump_suffix[256];
3738 char dump_suffix2[256];
3739 char dump_suffix3[256];
3740 struct obstack obst;
3743 ir_snprintf(problem_name, sizeof(problem_name), "%F_%s", chordal_env->irg, chordal_env->cls->name);
3744 ir_snprintf(dump_suffix, sizeof(dump_suffix), "-%s-remats", chordal_env->cls->name);
3745 ir_snprintf(dump_suffix2, sizeof(dump_suffix2), "-%s-pressure", chordal_env->cls->name);
3747 FIRM_DBG_REGISTER(si.dbg, "firm.be.ra.spillremat");
3748 DBG((si.dbg, LEVEL_1, "\n\n\t\t===== Processing %s =====\n\n", problem_name));
3750 obstack_init(&obst);
3751 si.chordal_env = chordal_env;
3753 si.cls = chordal_env->cls;
3754 si.lpp = new_lpp(problem_name, lpp_minimize);
3755 si.remat_info = new_set(cmp_remat_info, 4096);
3756 si.interferences = new_set(cmp_interference, 32);
3757 si.all_possible_remats = pset_new_ptr_default();
3758 si.spills = pset_new_ptr_default();
3759 si.inverse_ops = pset_new_ptr_default();
3760 #ifndef EXECFREQ_LOOPDEPH
3761 compute_execfreq(chordal_env->irg, LOOP_WEIGHT);
3763 si.execfreqs = NULL;
3768 si.n_regs = get_n_regs(&si);
3770 set_irg_link(chordal_env->irg, &si);
3771 compute_doms(chordal_env->irg);
3773 /* compute phi classes */
3774 // phi_class_compute(chordal_env->irg);
3776 be_analyze_regpressure(chordal_env, "-pre");
3778 #ifdef COLLECT_REMATS
3779 /* collect remats */
3780 DBG((si.dbg, LEVEL_1, "Collecting remats\n"));
3781 irg_walk_graph(chordal_env->irg, walker_remat_collector, NULL, &si);
3784 /* insert possible remats */
3785 DBG((si.dbg, LEVEL_1, "Inserting possible remats\n"));
3786 irg_block_walk_graph(chordal_env->irg, walker_remat_insertor, NULL, &si);
3787 DBG((si.dbg, LEVEL_2, " -> inserted %d possible remats\n", pset_count(si.all_possible_remats)));
3790 DBG((si.dbg, LEVEL_1, "Connecting remats with keep and dumping\n"));
3791 connect_all_remats_with_keep(&si);
3792 /* dump graph with inserted remats */
3793 dump_graph_with_remats(chordal_env->irg, dump_suffix);
3796 /* insert copies for phi arguments not in my regclass */
3797 irg_walk_graph(chordal_env->irg, walker_regclass_copy_insertor, NULL, &si);
3799 /* recompute liveness */
3800 DBG((si.dbg, LEVEL_1, "Recomputing liveness\n"));
3801 be_liveness(chordal_env->irg);
3805 DBG((si.dbg, LEVEL_1, "\tBuilding ILP\n"));
3806 DBG((si.dbg, LEVEL_2, "\t endwalker\n"));
3807 irg_block_walk_graph(chordal_env->irg, luke_endwalker, NULL, &si);
3809 DBG((si.dbg, LEVEL_2, "\t blockwalker\n"));
3810 irg_block_walk_graph(chordal_env->irg, luke_blockwalker, NULL, &si);
3812 #ifndef NO_MEMCOPIES
3813 DBG((si.dbg, LEVEL_2, "\t memcopyhandler\n"));
3814 memcopyhandler(&si);
3822 ir_snprintf(buf, sizeof(buf), "%s-spillremat.ilp", problem_name);
3823 if ((f = fopen(buf, "wt")) != NULL) {
3824 lpp_dump_plain(si.lpp, f);
3831 DBG((si.dbg, LEVEL_1, "\tSolving %F\n", chordal_env->irg));
3833 lpp_set_time_limit(si.lpp, ILP_TIMEOUT);
3837 lpp_solve_cplex(si.lpp);
3839 lpp_solve_net(si.lpp, LPP_SERVER, LPP_SOLVER);
3841 assert(lpp_is_sol_valid(si.lpp)
3842 && "solution of ILP must be valid");
3844 DBG((si.dbg, LEVEL_1, "\t%s: iterations: %d, solution time: %g, objective function: %g\n", problem_name, si.lpp->iterations, si.lpp->sol_time, is_zero(si.lpp->objval)?0.0:si.lpp->objval));
3846 #ifdef DUMP_SOLUTION
3851 ir_snprintf(buf, sizeof(buf), "%s-spillremat.sol", problem_name);
3852 if ((f = fopen(buf, "wt")) != NULL) {
3854 for (i = 0; i < si.lpp->var_next; ++i) {
3855 lpp_name_t *name = si.lpp->vars[i];
3856 fprintf(f, "%20s %4d %10f\n", name->name, name->nr, name->value);
3863 writeback_results(&si);
3867 kill_all_unused_values_in_schedule(&si);
3869 #if defined(KEEPALIVE_SPILLS) || defined(KEEPALIVE_RELOADS)
3870 be_dump(chordal_env->irg, "-spills-placed", dump_ir_block_graph);
3873 // move reloads upwards
3874 be_liveness(chordal_env->irg);
3875 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
3876 move_reloads_upward(&si);
3878 #ifndef NO_MEMCOPIES
3879 verify_phiclasses(&si);
3880 assign_spillslots(&si);
3883 irg_block_walk_graph(chordal_env->irg, walker_pressure_annotator, NULL, &si);
3885 dump_pressure_graph(&si, dump_suffix2);
3887 be_analyze_regpressure(chordal_env, "-post");
3889 free_dom(chordal_env->irg);
3890 del_set(si.interferences);
3891 del_pset(si.inverse_ops);
3892 del_pset(si.all_possible_remats);
3893 del_pset(si.spills);
3894 #ifndef EXECFREQ_LOOPDEPH
3898 obstack_free(&obst, NULL);
3899 DBG((si.dbg, LEVEL_1, "\tdone.\n"));
3902 #else /* WITH_ILP */
3905 only_that_you_can_compile_without_WITH_ILP_defined(void)
3909 #endif /* WITH_ILP */